Study On Density Sensitive Sparse Clustering Algorithm

Clustering is called an unsupervised learning in machine learning. Different fromclassification, class tag is not necessary to data object to cluster, and it need calculate byclustering learning. The data objects are divided into multiple classes, or clusters, thesimilarity between data objects in the same cluster or same class is very high, but the degreeof similarity between data objects in different clusters or classes is small.In the face of high-dimensional data, there are many deficiencies of the traditionalclustering algorithm. For example, the traditional clustering algorithm is often not accurate tohigh dimensional data sets with more complex distribution. As a result, people propose avariety of different ways to deal with high dimensional data sets. Feature selection anddimension reduction are used frequently. In this thesis, we use a Lasso method to control theweights of the data objects’ features, the smaller weights will be compressed to0automatically, so as to select the features of data objects, to reduce the feature dimensions,and to do a sparse clustering.The function of Euclidean distance or Manhattan distance is used in traditional clusteringalgorithm to represent the degree of similarity between data objects. Although the calculationamount of traditional methods is not heavy, and speed is fast, but the function value is not ableto accurately reflect the similarity of data objects with more complex distribution, and theclustering results will also be uncertain.In the face of this problem, this thesis adopts the density sensitive distance instead of thetraditional calculation method, and proposes the density sensitive sparse clustering algorithmbased on a model of sparse clustering. It’s faster and better to cluster to the data objects withmore complex distribution. The experimental results show that the density sensitive sparseclustering algorithm can cluster data objects of different distribution.